The Archive

datasette-llm 0.1a9 improves UX by auto-detecting default LLM models, reducing configuration overhead for Datasette Agent users.

Malware operators deploy LLMs to prioritize targets by financial value, escalating threat sophistication and attack targeting.

PSM-Bench evaluates LLM reasoning on perovskite photovoltaics; OpenPerov framework learns from 4,532 papers to guide materials design.

TopoQuant uses topological data analysis to measure structural drift in quantized LLMs, revealing compression-performance tradeoffs.

Log-Driven Semantic Tuning uses LLMs to automate MIP solver parameter tuning by interpreting performance logs.

LLM-based agent discovers Hamiltonian/Lindbladian models for quantum emitter disorder in color-center cavity-QED systems.

Trail framework enables local smart home control using typed decision models instead of LLM generation, eliminating cloud dependency.

User study with 14 physicians on MIMIC-IV finds multi-turn conversational text-to-SQL improves EHR query usability over single-turn.

Multi-estimate elicitation format fails to consistently debias LLM reasoning across anchoring, base-rate, and conjunction tasks.

Large-scale study (360M–3.1B params, 200B tokens) finds orthogonality constraints during LLM pre-training yield inconsistent improvements.

Total variation distance outperforms cross-entropy and KL divergence for training speculative decoding drafters to accelerate LLM inference.

Study evaluates LLM agents' ability to design therapeutic antibodies via task-guided and black-box protein design workflows.

SFTProbe accelerates supervised fine-tuning data generation for reasoning LLMs from weeks to hours via principled sampling.

BitLLM optimizes low-bit LLM inference on mobile GPUs by eliminating data-type conversion and memory bottlenecks.

White-box audit reveals low-margin token positions in greedy LLM decoding can redirect multi-agent deliberation.

Skill++ addresses routing interference in lifelong LLM agents where metadata-based skill selection breaks under new skills.

AC3 removes LLM-generated noise from multi-turn conversations via training-free context curation without retraining.

LLM agents exhibit social influence susceptibility in delegated audit tasks, mirroring human behavioral biases in judgment.

Benchmark finds five System-1 decision models statistically match GPT-5.6 judge performance on eleven tasks at lower cost without reasoning.

TAILOR framework uses Conditional Value at Risk to synthesize robust LLM heuristics, addressing tail-risk failures in average-case evals.

Large-scale audit of 12 System One decision models on BBQ, CrowS-Pairs, SilverSpoon benchmarks reveals widespread demographic bias in LLM typed-answer outputs.

LLMs generate coherent comedy improvisations but fail spontaneous dialogue features; analysis via linguistic measures and annotation.

VoirDire calibration method mitigates shared biases in LLM-judge panels via intervention-controlled surface-attribute evaluation.

Integer-exact MXFP4 kernel optimization runs 4-bit LLM weights on INT8 matrix units (Apple M5 GPUs) via single INT8 GEMM.

Breunig embedded a working NES 6502 emulator into Qwen3.5-4B weights, executing ROM code via LLM inference.

Flide inference engine accelerates LLM agent decision models via exact state forking, eliminating prefix recomputation for categorical reasoning workloads.

DecisionRank distills ranking knowledge from generative LLMs to decision models via on-policy distillation, reducing decoding latency.

ACCT enables KV cache reuse across different-scale LLMs in same family, reducing context processing overhead in heterogeneous deployments.

Zyphra introduces tree-structured hierarchical indexer achieving asymptotically subquadratic sparse attention for long-context LLMs.

Audit reveals persona-dependent variation in smoking-cessation health advice across six LLMs, raising fairness concerns.